 |
 |
Hi, each category has a different efficiency. If you correct by the efficiency before putting together the results you will get a larger error. For instance: suppose to have just two categories eff(1) = 90% eff(2) = 10% while eff(overall) = 50% (same amount of events in both categories at the origin) Suppose to measure N(1) = 900 +- 30 N(2) = 100 +- 10 => Then N(1)_origin = 1000 +- 33 N(2)_origin = 1000 +- 100 => Ntot_origin = 2000 +- 105 while using just one category and one efficiency you get N = 1000 +- 32 => (eff = 50%) Ntot_origin = 2000 +- 64 The effect depends on the difference in the efficiencies and on the the number of events in each category. Since in our categorization we have two "bad" categories (the last two, ch3ne1 and ch3ne2) with small efficiencies and containing a pretty large fraction of events at the origin, the final result can have a much different error. Daniele On Mon, 29 Apr 2002, Oliver Buchmueller wrote: > > Hi Daniele, > > thanks for the quick answer. This 20-30% difference was actually > the trigger for me to look more carefully to your results from the first > place. I just have difficulties to understand why a simple > (- assume statistically independent-) categorization can blow > up your fit error. > Looking at your numbers of fitted events and just adding the > the errors in quadrature I get 509+-44.3 events for the categorization > whereas you quote 559 +- 44 for no cat. . There is no 20-30% effect. > Are I am missing something (e.g. categories are not independent ..?) > > > Oliver > > > On Mon, 29 Apr 2002, Daniele del Re wrote: > > > > > Hi Oliver, > > > > thanks for your good comment. > > > > > ^ ^ > > > Are the two BR results | | > > > obtained from the same MC sample? > > > If yes, what has caused the shift? > > > > > > > As you see in general there is an increase of 20-30% in the error (due to > > categories with low statistics). This means that the two results can be > > different. > > > > In this particular case > > > > sqrt ( (sigma*1.3)^2 - sigma^2) ) ~ .8 sigma > > > > and 0.0179 - 0.0156 = .0023 = 1.6 * .0014(=sigma(ratio)) > > > > So this result is 1.6 sigma off. > > Looking in detail > > > > * signal events from the fit w/o categories: > > > > S = 559 +- 44 > > > > > > * signal events from the fit with categories: > > > > S S from truth > > > > 32 +- 7 115 > > 31 +- 12 93 > > 220 +- 20 662 > > 190 +- 28 594 > > 15 +- 15 108 > > 21 +- 19 144 > > > > total 509 1716 > > > > > > The lack comes from the last two categories and they weight more since > > they have a small efficiency. I don't see a fitting problem in these fits > > > > http://www.slac.stanford.edu/~daniele/vub/MCmulti/newMCshapech3ne1fitresults.eps > > http://www.slac.stanford.edu/~daniele/vub/MCmulti/newMCshapech3ne2fitresults.eps > > > > BTW I will look into the problem more in detail. > > > > Thanks a lot, > > > > Daniele > > > > > > > > > >